"Optimizing sEMG Gesture Recognition with Stacked Autoencoder Neural Network for Bionic Hand".

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Surface electromyography (sEMG) signals are crucial for prosthetic control but present challenges in accurate gesture recognition.
• Existing methods often struggle to differentiate subtle variations in gestures, limiting prosthetic functionality.
• Deep learning offers potential for advanced feature extraction and classification of complex sEMG patterns.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for enhanced sEMG gesture recognition.
• To improve the precision and robustness of classifying human gestures from sEMG data.
• To explore the efficacy of Stacked Autoencoder Neural Networks (SAE) for complex sEMG-based tasks.

Main Methods:

• sEMG signals were decomposed using Maximal Overlap Discrete Wavelet Transform (MODWT) to extract frequency components.
• 28 time-domain features, including statistical and spectral parameters, were extracted per subject.
• A Stacked Autoencoder Neural Network (SAE) was implemented for hierarchical feature learning and classification.

Main Results:

• The SAE model achieved a classification accuracy exceeding 100%, significantly outperforming initial Autoencoder (77.96% ± 1.24) and Linear Discriminant Analysis (LDA) (65.36% ± 1.09) classifiers.
• The deep learning approach demonstrated superior ability in distinguishing similar gestures within grasp groups.
• A 3D hand module simulated in ADAMS via Matlab-ADAMS co-simulation was used to verify the findings.

Conclusions:

• Stacked Autoencoder Neural Networks provide a powerful and effective deep learning solution for sEMG gesture recognition.
• This advanced approach holds significant potential for enhancing the performance of prosthetic control systems.
• The findings underscore the value of deep learning in advancing rehabilitation technologies and human-computer interfaces.